{
  "schema": "anvil-serving.os-comparison/v1",
  "rows": [
    {
      "context_target": 4096,
      "actual_prompt_tokens": 2969.0,
      "sample_count_per_os": 3,
      "metrics": {
        "decode_tok_s_p50": {
          "windows": 112.06564058179691,
          "linux": 149.01900227284116,
          "change_percent": 32.974747209937135
        },
        "effective_prefill_tok_s_p50": {
          "windows": 16729.465552552232,
          "linux": 23143.00551077051,
          "change_percent": 38.33678928995931
        },
        "ttft_p50_ms": {
          "windows": 177.4718000087887,
          "linux": 128.2897810015129,
          "change_percent": -27.712582508793083
        },
        "e2e_p50_ms": {
          "windows": 500.499099958688,
          "linux": 611.2129540015303,
          "change_percent": 22.120689937700334
        },
        "throughput_tok_s": {
          "windows": 56.996474897490415,
          "linux": 106.75009323273754,
          "change_percent": 87.29244821671911
        }
      },
      "windows_source": "windows-capacity-4k-r3.json",
      "linux_source": "linux-legacy-capacity-4k-r3.json",
      "limitations": [
        "Historical variable-output, shared-prefix workload; descriptive medians only",
        "Whole migrated stack comparison, not isolated OS causality"
      ],
      "median_output_tokens": {
        "windows": 36,
        "linux": 74
      }
    },
    {
      "context_target": 120000,
      "actual_prompt_tokens": 96278.0,
      "sample_count_per_os": 3,
      "metrics": {
        "decode_tok_s_p50": {
          "windows": 96.16500114296164,
          "linux": 125.02757587394592,
          "change_percent": 30.01359578634679
        },
        "effective_prefill_tok_s_p50": {
          "windows": 5748.638371968673,
          "linux": 5879.3029344226425,
          "change_percent": 2.272965422405271
        },
        "ttft_p50_ms": {
          "windows": 16747.966599999927,
          "linux": 16376.091381996957,
          "change_percent": -2.2204201076145624
        },
        "e2e_p50_ms": {
          "windows": 17233.929299982265,
          "linux": 16738.361063999037,
          "change_percent": -2.875538290526336
        },
        "throughput_tok_s": {
          "windows": 2.715436165889082,
          "linux": 2.6245311985354314,
          "change_percent": -3.3477114467128977
        }
      },
      "windows_source": "windows-capacity-120k-r3.json",
      "linux_source": "linux-legacy-capacity-120k-r3.json",
      "limitations": [
        "Historical variable-output, shared-prefix workload; descriptive medians only",
        "Whole migrated stack comparison, not isolated OS causality"
      ],
      "median_output_tokens": {
        "windows": 48,
        "linux": 46
      }
    },
    {
      "context_target": 262144,
      "actual_prompt_tokens": 216101.0,
      "sample_count_per_os": 3,
      "metrics": {
        "decode_tok_s_p50": {
          "windows": 102.41880883554646,
          "linux": 124.46274256962307,
          "change_percent": 21.52332563198669
        },
        "effective_prefill_tok_s_p50": {
          "windows": 5578.783369064824,
          "linux": 5458.399114754668,
          "change_percent": -2.1578944071874995
        },
        "ttft_p50_ms": {
          "windows": 38736.223999992944,
          "linux": 39590.5463229974,
          "change_percent": 2.2054868409595363
        },
        "e2e_p50_ms": {
          "windows": 39107.24919999484,
          "linux": 39967.26827999737,
          "change_percent": 2.1991295670129807
        },
        "throughput_tok_s": {
          "windows": 1.2242969681044664,
          "linux": 1.0338363102010584,
          "change_percent": -15.556736875554888
        }
      },
      "windows_source": "windows-capacity-262k-r3.json",
      "linux_source": "linux-legacy-capacity-262k-r3.json",
      "limitations": [
        "Historical variable-output, shared-prefix workload; descriptive medians only",
        "Whole migrated stack comparison, not isolated OS causality"
      ],
      "median_output_tokens": {
        "windows": 39,
        "linux": 48
      }
    },
    {
      "context_target": 380000,
      "actual_prompt_tokens": 304491.0,
      "sample_count_per_os": 3,
      "metrics": {
        "decode_tok_s_p50": {
          "windows": 99.79115138353907,
          "linux": 120.28614187074675,
          "change_percent": 20.53788357290003
        },
        "effective_prefill_tok_s_p50": {
          "windows": 5456.732792727141,
          "linux": 5425.364604799794,
          "change_percent": -0.5748529224145926
        },
        "ttft_p50_ms": {
          "windows": 55801.15680000745,
          "linux": 56123.41785300305,
          "change_percent": 0.5775167962029704
        },
        "e2e_p50_ms": {
          "windows": 56362.32839996228,
          "linux": 56462.859640003444,
          "change_percent": 0.1783660166197576
        },
        "throughput_tok_s": {
          "windows": 0.8353306078293593,
          "linux": 0.7631981612491053,
          "change_percent": -8.635197358288249
        }
      },
      "windows_source": "windows-capacity-380k-r3.json",
      "linux_source": "linux-legacy-capacity-380k-r3.json",
      "limitations": [
        "Historical variable-output, shared-prefix workload; descriptive medians only",
        "Whole migrated stack comparison, not isolated OS causality"
      ],
      "median_output_tokens": {
        "windows": 51,
        "linux": 48
      }
    }
  ],
  "endurance": {
    "sample_count_per_os": 60,
    "windows_source": "windows-endurance-4k-r60.json",
    "linux_source": "endurance-4k-r60.json",
    "decode_windows": 102.19275206464441,
    "decode_linux": 142.6429205479183,
    "decode_change_percent": 39.58222835381338,
    "limitations": [
      "Variable-output shared-prefix historical workload",
      "n60 p99 is descriptive only"
    ]
  }
}